Every AI content tool can generate a sentence that ends in a footnote. Far fewer can promise that the footnote points somewhere real. That distinction is the premise Citeya is built on, and 'source-backed' only means something if we're precise about which checks are automated and which aren't. This post walks through what actually happens between a topic and a cited draft, and where a human still has to look.
Step One: Sources Come Before the Draft
Citeya gathers sources before a single sentence is written. It searches recent news coverage for your title, keywords, and category, and if that search comes back empty it falls back to a web search that asks specifically for recent, credible pages with real URLs. You can also supply your own sources. Either way, the draft is built around the sources rather than having citations patched onto finished text, which is the failure mode behind plausible-looking but wrong citations. Peer-reviewed testing of ChatGPT's bibliographic citations found exactly that: references that look right and turn out to be fabricated or wrong.
Step Two: Every Link Has to Actually Resolve
Before a source can be cited, Citeya requests its URL. Anything that returns 'not found' (a 404 or 410) or can't be reached at all is dropped, so a stale search result, a deleted page, or a mis-parsed link never becomes a citation. Some major publishers answer every automated request with a 401, 403, or 429; those are treated as live, because the server is clearly real and responding, rather than silently discarding well-known sources.
Step Three: Primary Sources First
The sources that survive the link check are then classified, from their titles and snippets, as primary (original research, official data, or the organisation's own statement), secondary reporting, or an aggregator summary, and older items are flagged when the topic moves fast. Primary sources go to the top of the list the draft is written from, so a statistic is more likely to be cited to its origin than to a copy. This is a model's judgment from limited text, not a guarantee, which is why the reviewer still checks.
Step Four: Citation Rules in the Drafting Instructions
The drafting instructions require the article to cite the provided sources inline as real links, to copy every URL character for character from the source list, and never to invent a quote, a statistic, or a source to hit a target. If no sources survive the earlier steps, the draft is told not to cite anything rather than fabricate an attribution. Each category also carries its own authority requirement, such as a clinical finding or named institution for health topics, or a standards body or official documentation for technical ones.
Step Five: A Quality Gate, Then a Claim Check
Each draft is scored on areas including SEO, authority, user experience, and readability. If it falls below the bar, it's regenerated once with the specific gaps fed back in, such as too few attributed statistics or a missing citation, rather than a vague 'try again'. The gate can't make a weak source strong, but it does catch drafts that ignored the sources they were given. After that, every sentence that carries a citation is compared with the source it links to, and any claim the source doesn't clearly back (a different number, a missing date, a broader conclusion than the source draws) is listed in the article's quality report as a claim to verify. Nothing is rewritten automatically; the flags tell the reviewer where to look first. For automated campaigns on health, finance, or legal topics, the article also waits for an editor's approval before it's published.
What Automation Doesn't Decide: Is This the Right Source?
A live link isn't the same as a good source. The ordering and claim check above work from titles and short snippets, so they can't confirm that a source is the true original, that its full text says what the snippet suggests, or that a second independent outlet agrees. Those are judgment calls, and they're the reviewer's job. The CRAAP test, developed at CSU Chico's Meriam Library, is a good checklist for them: Currency, Relevance, Authority, Accuracy, and Purpose.
In practice that means a preference order when you review: an original study or dataset beats a news article summarizing it, which beats a blog post summarizing the news article. Each layer of removal is a place where a number can get rounded, a caveat dropped, or a finding overstated. If a draft cites the copy, swap in the original.
Currency and Corroboration Are Still Human Checks
A source can be authoritative and still stale. In fast-moving areas like AI tooling, search, and platform policy, check the date against how fast the topic moves; a five-year-old source is fine for a historical fact and misleading for how detection tools perform today. IFLA's guidance on spotting misinformation puts 'check the date' alongside checking the author and supporting sources.
For any claim that carries the argument, look for a second, independent source. The International Fact-Checking Network's Code of Principles asks fact-checkers to be transparent about their sources and methods for the same reason: a claim that appears in only one place deserves a second look before it's published as settled.
What a Reviewer Is Actually Checking, Beyond Whether the Link Works
The final human pass is looking for something narrower and easier to miss than 'is this source generally credible': does this specific source actually support the specific claim it's attached to, in this specific sentence. It's common for a source to be entirely legitimate and still be a mismatch, cited for a broader point than it makes, or attached to a number that's adjacent to what the source says rather than identical to it. Automation can confirm a link is live; it can't confirm that the source's conclusion and the draft's sentence make the same claim. That gap is exactly where a reviewer earns their keep.
The Honest Limits of This
These steps remove the most embarrassing failures: dead links, invented URLs, drafts that ignore their sources, and citations that plainly don't match their claim. They don't guarantee that every citation is the best available source, that it's current enough, or that it says exactly what the sentence claims. Primary sources get paywalled, niche topics surface weaker coverage, and a claim that's wrong everywhere because everyone copied the same original error will look well supported to any automated check.
That's why every Citeya draft is meant to be reviewed before it's published. Google's guidance on helpful, reliable, people-first content puts accuracy and trustworthy sourcing ahead of anything that merely looks well optimized, and that standard is met by the person who signs off, not by the tool.
A citation is only as good as the process that produced it. The bar isn't 'does this look like a source' — it's 'would this hold up if someone actually clicked it.'